Human And Machine Perception: Communication, Interaction, And Integration 9789812703095, 9789812384317

The theme of this book on human and machine perception is communication, interaction, and integration. For each basic to

170 87 18MB

English Pages 183 Year 2005

Report DMCA / Copyright

DOWNLOAD FILE

Polecaj historie

Human And Machine Perception: Communication, Interaction, And Integration
 9789812703095, 9789812384317

Citation preview

HUMAN &. MACHINE

PERCEPTION Communication, Interaction, and Integration

HUMAN & MACHINE

PERCEPTION Communication, Interaction, and Integration Santa Caterina di Pittinuri, Oristano, Italy September 6-9,2004

EDITORS

SERGIO V I T U L A N D Dipartimento di Scienze Mediche e Internistiche, Universita di Cagliari Policlinico Universitario, Monserrato, 09042 Cagliari, Italy

V I T D DI GESU Dipartimento di Matematica ed Applicazioni, Universita Di Palermo ViaArchirafi 34,90123 Palermo, Italy

V I R G I N I O CANTDNI * RDBERTD M A R M D - A L E S S A N D R A S E T T * Dipartimento Informatica, e Sistemistica, Universita di Pavia, via Ferrata I 27100 Pavia, Italy

\jjJ5 World Scientific NEW JERSEY

• LONDON • SINGAPORE • BEIJING

• SHANGHAI

• HONGKONG

• TAIPEI • CHENNAI

Published by World Scientific Publishing Co. Pte. Ltd. 5 Toh Tuck Link, Singapore 596224 USA office: 27 Warren Street, Suite 401-402, Hackensack, NJ 07601 UK office: 57 Shelton Street, Covent Garden, London WC2H 9HE

British Library Cataloguing-in-Publication Data A catalogue record for this book is available from the British Library.

HUMAN AND MACHINE PERCEPTION: COMMUNICATION, INTERACTION, AND INTEGRATION Copyright © 2005 by World Scientific Publishing Co. Pte. Ltd. All rights reserved. This book, or parts thereof, may not be reproduced in any form or by any means, electronic or mechanical, including photocopying, recording or any information storage and retrieval system now known or to be invented, without written permission from the Publisher.

For photocopying of material in this volume, please pay a copying fee through the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to photocopy is not required from the publisher.

ISBN 981-238-431-6

Printed in Singapore by World Scientific Printers (S) Pte Ltd

V

PREFACE

The following are the proceedings of the Sixth International Workshop on Human and Machine Perception held in Santa Caterina di Pittinuri (Oristano), Italy, on September 06-09, 2004, under the auspices of the following Institutions: the Italian Ministry of the University and the Researches, the University of Cagliari, the Pavia University, and the Inter-Department Centers of Cognitive Sciences of Palermo University. A broad spectrum of topics are covered in this series, ranging from computer perception to psychology and physiology of perception. The theme of this workshop on Human and Machine Perception was focused on Communication, Interaction and Integration. As in the past editions the final goal has been the analysis and the comparison of biological and artificial solutions. The focus of the lectures has been on presenting the state-of-the-art and outlining open questions. In particular, they sought to stress links, suggesting possible synergies between the different cultural areas. The panel discussion has been conceived as a forum for an open debate, briefly introduced by each panelist, and mainly aimed at deeper investigation of the different approaches to perception and strictly related topics. The panelists were asked to prepare a few statements on hot-points as a guide for discussion. These statements were delivered to the participants together with the final program, for a more qualified discussion. The number of participants to the workshop was limited to 70. Besides the 8 invited lecturers and panelists, more participants were admitted. Priority for these positions was given to young researchers who made significant contributions to the open discussions. Both the lectures and the contents of the panel discussions are included in these proceedings. The workshop structure consisted of three modules each of one organized in four general talks. The first one, Representing and coding for communication, was grounded on general talks and a panel discussion to give the foundations. The second one, Exploration, visualization and discovering in very large data set, was focused on methodologies for data mining, classification and recognition. The third one, Information exchange: Machine versus Machine, was dedicated to the communication and interaction of hybrid complex systems. In this edition we included, for the second time in the series, also a few solicited presentations from the audience.

VI

Preface

The description of both natural and artificial structures and basic approaches represents a natural starting point for a workshop which intends to analyze the processes of systems communication, interaction and integration from different viewpoints. The lectures were organized by alternating the functional descriptions of natural and artificial information management and decision making related to perception, learning, knowledge acquisition and storage. Further inquiries concern the human ability to exploit spatial information for reasoning and communication purposes and the new technologies and facilities on this field. Representing and coding for communication intends to addressed and to explore basic references for communication processes that allow natural or/and artificial systems to exchange knowledge and information to achieve a smart situated behaviour. Tools and modalities (languages, protocols, coding schemes, ect.) for handling perceptual information are discussed and compared. Advances in technology lead towards communications that are often mediated by computers and networks. Exploration and discovering in very large dataset explores how multimedia technology supplies a new way of gaining understanding into large data structures, allowing the user to gain insight into the process under investigation. Data-mining provides a new tool consisting of automatic discovering of patterns, changes, correlations and anomalies through an interactive display and analysis of data up to the elicitation of information and knowledge. Progress in statistics and data analysis are also shown in relation with data retrieving and discovering. Information exchange: machine versus machine is addressed to information exchange among the living and how human being and machines may interact for improving performances and/or behavior. In particular, the usability of a given multi-media interface, in terms of performance interface evaluation, is investigated. Moreover, how machines exchange information, learn by experimenting, and improve their performances is discussed. Modalities of interaction among distributed and heterogeneous systems, coordination in multiagent systems, control strategies and suitable hierarchies are compared. The workshop dealt with most of these problems and the results of the presentations and discussions are herewith included even if the chapters of the book vary somewhat from the scheduled program, to take care of the positions that emerged from the debates and to the audience reactions.

Preface

vn

Acknowledgements The workshop, and thus indirectly this book, was made possible through the generous financial support of the universities involved, and the research organizations that are listed separately. Their support is gratefully acknowledged. In particular we would like to thank Roberto Marmo and Alessandra Setti for the completion of this volume. We are also grateful for the assistance offered by World Scientific for the preparation of this volume.

Virginio Cantoni, Vito Di Gesu, Sergio Vitulano

Vlll

SPONSORING INSTITUTIONS

The following institutions are gratefully acknowledged for their contributions and support to the Workshop: • MURST, Ministero dell'Universita e della Ricerca Scientifica e Tecnologica, Italy • Dottorato di Ricerca in Tecnologie Biomediche Applicate alle Scienze odontostomatologiche, Seconda Universita di Napoli, Italy • Dottorato di ricerca in Alimenti e Salute: Biotecnologie e Metodologie Applicate alia Fisiopatologia Digestiva, Seconda Universita di Napoli, Italy • Universita di Cagliari, Italy • Universita di Pavia, Italy • Universita di Palermo, Italy • Provincia di Cagliari, Italy

IX

CONTENTS

Preface

v

REPRESENTING AND CODING FOR COMMUNICATION One Eye, Two Eyes ... How Many Eyes? Notes and Reflections on Vision Systems for Mobile Robotics G. Adorni, S. Cagnoni, and M. Mordonini A Note on Pictorial Information V. Di Gesit Biomedical Signal Processing and Modeling: Multiscale and Multiorgan Integration of Information S. Cerutti The Binding Problem and the Emergence of Perceptual Meaning E. Pessa

1

23

35

49

EXPLORATION AND DISCOVERING IN VERY LARGE DATASET Sequence Rule Models for Web Usage Mining P. Cerchiello and P. Giudici

71

An Improved 3D Face Recognition Method Based on Normal Map A. F. Abate, M. Nappi, S. Ricciardi, and G. Sabatino

77

Entropy Measures in Image Classification A. Casanova, V. Di Gesii, G. Lo Bosco, and S. Vitulano

89

Complex Objects Classified by Morphological Shape Analysis and Elliptical Fourier Descriptors B. Ballard, D. Tegolo, C. Volenti, and C. Tripodo

105

x

Contents

INFORMATION EXCHANGE: MACHINE VERSUS MACHINE Visual Attention Mechanisms for Information Visualization R. Marmo and M. Valle Integrating Human and Machine Perception to Reverse-Engineer the Human Vision System R. Piroddi and M. Petrou

Ill

119

Situated Vision: A Step Further Towards Autonomous Systems B. Zavidovique and R. Reynaud

131

Recognizing Occluded Face Using Fractals A. F. Abate, M. Nappi, D. Riccio, and M. Tucci

161

1

ONE EYE, TWO EYES... HOW MANY EYES? NOTES AND REFLECTIONS ON VISION SYSTEMS FOR MOBILE ROBOTICS* GIOVANNI ADORNI Dipartimento

di Informatica, Sistemistica e Telematica, Universita di Genova, Via all'Opera Pia 13, 16145 Genoa, Italy E-mail: [email protected]

STEFANO CAGNONI, MONICA MORDONINI Dipartimento

di Ingegneria dell'Informazione, Universita di Parma, Parco Area delle Scienze 181a, 43100 Parma, Italy E-mail: {cagnoni, monicaj @ce. unipr. it

In this chapter we discuss some issues on mobile robot perception from the point of view of the structure of vision systems. Through some experiences and results from research projects we will approach the problem from the point of view of purposive vision, with particular regard to less conventional sensors and systems, such as omnidirectional and hybrid systems.

1. Introduction Differentiation of living beings is the most outstanding result of million of years of natural evolution. The main characterizing feature of Darwinian evolutionary processes is having the fitness of an individual, with respect to the surrounding environment and to performing the actions required to survive in it, as both their goal and driving force. As a result, each species has developed its own specialized apparatus, which maximize performance in its specific environment. Sensory and locomotory systems are probably the ones for which species differentiation is most immediately evident.

This work was partially supported by ASI (Italian Space Agency) within the grants "Coordination of the cluster Robotic Vision" and "Hybrid Vision System for Long Range Rovering", and by ENEA (National Institute for Alternative Energy) within the grant "Sensori Intelligenti".

2

G. Adorni, S. Cagnoni, and M. Mordonini

For many living beings, among sensory systems, the vision system is the richest source of information which is used to realize where they are and what is around them, in order to plan what their immediate goal should be and what sequence of actions should be performed to reach it. A basic taxonomy of vision systems in animal species can be based, firstly, on the number of sensitive elements by which an eye is composed and on the number of eyes; secondly, on features such as their field of view and sensitivity to light or colour. Virtually all vertebrates, along with some invertebrates such as molluscs, medusas and worms, have simple eyes, similar to humans'. Obviously, despite their similar physical structure, their performance generally differs, according to the environmental conditions in which they are expected to operate. For instance, cats, bats, owls, which usually act in dark environments, if not exclusively at night-time, have only rods, more sensitive and numerous than in human eyes, to obtain high sensitivity to light. In dolphins' eyes, as well, there are 7000 times as many rods as in humans, to cope with limited underwater lighting. Fish eyes have usually a flat cornea and a spherical, non-deformable crystalline lens, to improve sight in the near field of view. On the other hand, birds' eyes are elongated in antero-posterior direction to produce larger images of distant objects. Among invertebrates, one of arthropods' peculiar features is having composite eyes, made up of many (up to thousands) simple eyes. The final image which they perceive is formed as apposition or superposition of the single images which are perceived by the simple eyes. Similarly to what has been happening in natural evolution through million of years, the choices made by designers of vision systems for mobile robots are mainly influenced by the tasks which the robots, on which such systems are going to be mounted, are expected to perform. In computer vision, the purposive vision approach aims at solving problems encountered in dealing with complex tasks or environments by identifying the goal of the task; this simplifies design by making explicit just that piece of information that is needed to achieve such a goal. As a matter of fact, mobile robotics is probably the field in which the purposive vision paradigm has been applied most frequently. Mobile robotics applications actually require that task-specific information be selected from the usually huge amount available, to satisfy the constraints of real-time or just-intime processing imposed by the dynamic environments in which mobile robotics tasks are usually performed. In designing robot vision systems the main degrees of freedom available to the designer are the same peculiarities that distinguish biological vision systems, namely the number of sensors and some of their features, of which the width of

Notes and Reflections on Vision Systems for Mobile Robotics

3

the field of view and the related parameters of distortion and resolution are perhaps the most relevant. Therefore, also in the case of robotic vision systems, a similar taxonomy to the one reported above for natural vision systems can be applied. In this chapter, we aim at discussing the task of designing robot vision systems according to such a taxonomy and from the point of view of purposive vision applications. The chapter reviews examples of applications or preliminary experiences we developed in previous research projects, highlighting the taskoriented motivation of design, and discussing the results that such systems can achieve. We will consider the case of single-sensor systems, as opposed to multisensor ones. As regards sensor types, we will consider traditional pin-hole and omnidirectional sensors in single or multi-sensor configurations. Particular attention will be given to the design of omnidirectional and hybrid omnidirectional/pin-hole systems. 2. Single-sensor Systems Up to a few years ago, most computer vision applications relied on single-sensor systems. However, the choice very often did not depend on an objective evaluation of the specifications on the design of the vision system, but mostly on constraints imposed by the computation power that was available. As processor power increases, along with sensor technology, it is possible to design computer vision applications which are more and more complex, both intrinsically and from the point of view of the sensors by which information is acquired. Mobile robotics applications, in particular, have become more and more reliant on heterogeneous sensory systems (cameras, ultrasonic sensors, laser scanners, etc.), in which even the vision system is often made up of multiple, homogeneous or heterogeneous, components. In the far-from-complete review of vision sensors made by this chapter, attention is principally focused onto multi-sensor vision systems. Therefore, discussion of single-sensor system will be limited to describing their essential features, considering them mainly as basic components of more complex systems. Also, since the main focus of this chapter is on less conventional sensor design, in describing conventional cameras we will limit our discussion to a very concise review of the main features which affect performance of mobile robot vision systems. In particular we will focus on the features which distinguish conventional from omnidirectional sensors, discussing the latter more amply, even if, also in their case, with the aim of introducing them as components of multi-sensor systems.

4

G. Adorni, S. Cagnoni, and M. Mordonini

2.1. Conventional Cameras Conventional cameras use the classic pin-hole principle, by which objects in a scene are projected onto a planar sensor after being focused on it by a lens, or simply after the light rays which they reflect pass through a tiny hole. If the size and resolution of the sensor are set, the field of view of such a system is inversely dependent on the focal distance, i.e. the distance between the lens (hole) and the sensor onto which the image is projected. Conversely, the resolution of the image which can be obtained increases with the focal distance, since a smaller scene is projected onto the same sensitive area. Images acquired by conventional cameras are affected by two kinds of distortions: perspective effects and deformations that derive from the shape of the lens through which the scene is observed. Both kinds of distortion hamper direct measures, since distances in pixel are not proportional to actual distances in the real world. In mobile robotics applications, being able to reconstruct the surrounding scene with high accuracy is a critical task in navigation, especially as obstacle detection and motion planning are concerned. After introducing omnidirectional sensors in the next section, we introduce the Inverse Perspective Transform (IPT) as a general tool to, firstly, compensate for the classical perspective distortion in pin-hole images and, more generally, to compensate for any measurable distortion in images acquired by any sort of vision sensor. IPT can be effectively used to detect obstacles in mobile robotics application based on stereo vision systems, as shown in Section 3.2. 2.2. Omnidirectional Cameras Being able to see behind one's shoulders has probably been one of the most desirable supernatural faculties for humans (possibly the most desirable after being able to fly), resulting in several mythological creatures, such as Janus, the divinity with two faces, and Argus Panoptes, the giant with a hundred eyes. In applications involving dynamic environments such as mobile robotics, widening the field of view is of major importance, to be able to enhance awareness about what is happening in the surroundings and to obviate the need for active cameras, that require complex control strategies. In this section we will consider a particular class of omnidirectional sensors, catadioptric sensors, which consist of a conventional camera that acquires the image reflected on a convex mirror. Catadioptric omnidirectional sensors suffer from two main limitations. The most relevant one is that the near field, which is the least distorted part of the

Notes and Reflections on Vision Systems for Mobile Robotics

5

image, is partially obstructed by the reflection of the camera on the mirror. A further limitation is that the accumulation of camera and system distortions makes it quite difficult either to find the resulting distortion law and to compensate for it, or to design a mirror profile that can achieve a good trade-off between width of the field of view, image resolution and distortion (see [5] for a discussion of the problem). To reduce this effect, mirrors that produce a nondistorted image of a reference plane have been recently described [13,16]. Omnidirectional systems are therefore very efficient as concerns detection of target position, but critical from the point of view of the accuracy with which the target is detected. For these reasons, either the omnidirectional vision sensor has been integrated with a different kind of sensor, to make object detection and robot self-localization more precise and robust (see, for example [9,12,15,19], or arrays of omnidirectional sensors have been used to implement triangulation algorithms [20]. In the following section we discuss the problem of designing omnidirectional systems from the point of view of mobile robotics applications. 2.2.1. Design of catadioptric omnidirectional systems One of the most important issues in designing a catadioptric omnidirectional vision system is choosing the geometry of the reflecting surface of the mirror [14]. Different mirror profiles have been proposed in literature [21]. In particular, adopting the solution proposed by [13] one can design a mirror which preserves the geometry of a plane perpendicular to the axis of symmetry of the reflecting surface, i.e., an isometric mirror: the mirror acts as a computational sensor, capable of providing distortion-free images automatically, thus eliminating the need for further processing, if the distortion caused by the camera lens is neglected. Other researchers propose the use of multi-part mirrors [8], composed of a spherical mirror providing better resolution in the proximity of the robot to accurately localize significant features in the environment, merged with a conic mirror that extends the field of view of the system, to perceive objects at a greater distance at the price of a loss in resolution. This composite mirror shape permits to image objects that are closer to the sensor with respect to classical off-the-shelf conical mirrors that are often used to achieve omnidirectional vision. This design has been suggested by the requirements of RoboCup, in which robots need to see the ball up to the point at which they touch it.

Visit http://www.robocup.org for information on the RoboCup robot soccer competition.

6

G. Adorni, S. Cagnoni, and M. Mordonini

Obviously, to design the profile of a mirror it is always necessary to first define the requirements for the particular application for which the catadioptric vision system is meant; thus a general rule for generating the optimum mirror profile does not exist. For example, a vision system which is meant only for obstacle avoidance probably does not require that the robot recognize very distant objects; the opposite is true for a localization system which relies on the recognition of significant features distributed on the walls or extracted from the surrounding environment [9]. Once the requirements for a particular application have been specified, to design a mirror profile with the desired characteristics, a differential equation needs to be solved, which can be inferred by applying the laws of linear optics. The following specifications, suggested by considerations related with both the optics and the mechanics of the system, are to be met in mobile robot vision systems: 1. minimizing encumbrance and weight of the mirror by minimizing its radius. If one considers that the whole vision system is typically made up of a camera and a transparent cylinder holding the mirror, by reducing the mirror weight (and, if possible, the cylinder height) the mechanical stability of the whole system is increased. Notice also that, for a given focal length of the camera, decreasing the radius of the mirror also requires that the distance between the camera and the reflecting surface be decreased as well as, consequently, the height of the support; 2. designing profiles which reduce the area of the image containing useless information, e.g., the reflection of the robot on the mirror; 3. radially extending the field of view up to where relevant information can be still extracted; 4. keeping the resolution of the most relevant part of the image as high as possible, to provide the clearest possible view of the objects that lie close to it and require that the robot take strictly real-time decisions in dealing with them. The rationale behind these basic requirements is clear if one observes Figure 1, in which a cheap but clearly non-optimal catadioptric surface (the convex face of a ladle!) has been used to acquire an omnidirectional view of a RoboCup field, in which a ball is placed at a distance of 1 m from the sensor. In such an image, the most evident problems are: 1. poor resolution already at a relatively short distance: the ball is hardly visible, despite being at a distance of about 1 meter; 2. extension of the field of view beyond the area of interest: useless lowresolution information (very distant objects) is visible; 3. the image region occupied by the reflection of the robot body on the mirror is very large.

Notes and Reflections on Vision Systems for Mobile Robotics

7

;;

' ™

Figure 1. A 'home-made' omnidirectional sensor obtained using a kitchen ladle.

According to the above-mentioned guidelines, we have designed a mirror which is very similar to the one in Figure 2. The mirror is obtained as a revolution surface by revolving the profile around the z~axis, it has a radius of 5 cm and it is about 2.69 cm high. The profile can be decomposed in two sections: the first section (from point A to point B) can be modeled as an arc of circumference, while the second section (from point B to point C) is a straight line. The tangent to the circumference in point B has the same slope as the straight line, preserving the continuity in the first order derivative. Observing the resulting surface (see Figure 3), it is straightforward to see that it is composed by a conic mirror jointed to an "almost spherical" mirror. On one hand, the property of the conic part to reflect objects which are far from the robot is very useful to detect objects of interest (other robots, openings, signs on walls, etc.). Obviously, resolution decreases with object distance from the robot, allowing only for an approximate perception of the distance of interesting features. However, if focusing on a particular feature

Figure 2. Catadioptric omnidirectional sensor for mobile robotics applications and its generating profile.

G. Adomi, S. Cagnoni, and M. Mordonini

8

is needed, the system provides sufficient information for the robot to move towards it, until it can perceive its geometric properties (position and shape) more accurately. On the other hand, the property of the spherical part to reflect close objects with a higher resolution is well suited for obstacle detection and avoidance. As regards reducing the extension of the image region containing useless information, since this area is situated in the center of the image and corresponds to the reflection of the camera onto the central part of the mirror, one can introduce a discontinuity in the gradient of the revolution surface in correspondence of point (x=0, y=0, z=0). In Figure 3 it can be easily seen that, since the first order derivative of the curve in correspondence of (x=0, z=0) is not equal to zero, the resulting revolution surface will have a discontinuity in (x=0, y=0, z=0). By varying the slope of the tangent in such a point one can obtain different revolution surfaces. This means that a distortion is introduced in the reflected image, such that the central area which contains useless information is significantly shrunk.

3!Hr- r *

A-^y^2 5 m

Figure 3. The resulting views (below) of a virtual environment (above) as taken from a mirror with no discontinuities (left) and a mirror with a discontinuous profile in the point (x=0, y=0,z=O) (right).

Notes and Reflections on Vision Systems for Mobile Robotics

9

2.2.2. Inverse Perspective Transform While distortions due to lens shape can be completely removed if the lens model and its corresponding parameters are known or if a proper empirical calibration using a reference shape is performed, perspective effect removal from a single view is possible only with respect to an arbitrary chosen plane. Everything that lies on such a plane will be reconstructed as seen in an orthogonal projection onto the plane, i.e., with no perspective distortion. The problem can therefore be formalized as that of finding a function Px>y= C(Iy) that maps each pixel in the image Iy onto the corresponding point P x y of a new image P (with coordinates x,y) that represents an orthogonal projection onto the reference plane. Therefore, limiting one's interest to objects lying on the reference plane, it is possible to reason on the scene observing it with no distortions, at least on the plane level. The most appealing feature, in this case, is that a direct-proportionality relationship between distances in the reconstructed image and in the real world can be obtained, which is a fundamental requirement for geometrical reasoning. This transformation is often referred to as Inverse Perspective Transform (IPT) [1,17,18], since perspective-effect removal is the most common aim with which it is performed, even if it actually represents only one of the problems for which it provides a solution. For any different plane, one such function must be computed that differs from any other. Everything that does not lie on the reference plane is affected by a distortion that depends on its distance from the plane and on its position with respect to the camera. If all parameters related to the geometry of the acquisition systems and to the distortions introduced by the camera were known, the derivation of C could be straightforward. However, this is not always the case, most often because of the lack of an exact model of camera distortion. However, it is often possible to effectively (and efficiently) derive C empirically using proper calibration algorithms. An empirical derivation of C through a point-to-point mapping between the actual view and the transformed one is virtually possible for any kind of cameras. We will consider here the problem of computing C0, the generalization of the IPT for a catadioptric omnidirectional sensor. In this case, the problem is complicated by the non-planar profile of the mirror; on the other hand, the circular symmetry of the device provides the opportunity of dramatically simplifying such a procedure. If the reflecting surface were perfectly manufactured, it would be sufficient to compute just the restriction of C0 along one radius of the mirror projection on the image plane to compute the whole function. However, possible manufacturing flaws may affect both shape and surface smoothness of the

10

G. Adorni, S. Cagnoni, and M. Mordonini

mirror. In addition to singularities that do not affect sensor symmetry and can be included in the radial model of the mirror (caused, for example, by the joint between two differently shaped surfaces, as discussed before), a few other minor isolated flaws can be found scattered over the surface. Similar considerations can be made regarding the lens through which the image reflected on the mirror is captured by the camera. To account for all sorts of distortions an empirical derivation of C0 based on an appropriate sampling of the function in the image space can be made. Choosing such a procedure to compute C0 permits to include also the lens model into the mapping function. The basic principle by which C0 can be derived empirically is to consider a set of equally-spaced radii, along each of which values of C0 are computed for a set of uniformly-sampled points for which the relative position with respect to the sensor is known exactly. This produces a polar grid of points for which the values of C0 are known. To compute the function for a generic point P located anywhere in the field of view of the sensor, an interpolation can be made between the set of points among which P is located. The number of data-points (interpolation nodes) needed to achieve sufficient accuracy depends mainly on the mirror profile and on the mirror surface quality. This calibration process can be automated, especially in the presence of well manufactured mirrors, by automatically detecting relevant points. To do so, a simple pattern consisting of a white stripe with a set of aligned black squares superimposed on it can be used, as shown in Figure 4. The reference data-points, to be used as nodes for the grid, are extracted by automatically detecting the squares in a set of one or more images grabbed turning the robot around the vertical axis of the sensor. Doing so the reference pattern is reflected by different mirror portions in each image.

Figure 4. An example of IPT transformation of an image acquired by an omnidirectional camera.

Notes and Reflections on Vision Systems for Mobile Robotics

11

If distances between the shapes forming the pattern are known exactly, the only requirement is that one of the shapes, at known distance, be distinguishable (e.g., by its colour) from the others. The shape should be possibly located within the highest-resolution area of the sensor. This makes it possible to use the reference shape as a landmark to automatically measure the distance from the camera of every shape on the reference plane, while removing the need to accurately position the robot at a predefined distance from the pattern, which could be a further source of calibration errors. Operatively, in the first step of the automatic calibration process, the white stripe, along with the centres of the reference shape, are easily detected. These reference points are inserted into the set of samples on which interpolation is then performed. The process can be repeated for different headings of the robot, simply turning the robot around its central symmetry axis. In the second step, interpolation is performed to compute the function C0 from the point set extracted as described. A look-up table that associates each pair of coordinates in the IPT-transformed image to a pair of coordinates in the original image can thus be computed. This calibration process is fast and provides good results, as shown in Figure 4. The IPT plays an important role in several applications in which finding a relevant reference plane is easy. This is true for most indoor Mobile Service Robotics applications (such as surveillance of banks and warehouses, transportation of goods, escort for people at exhibitions and museums, etc.), since most objects which the robot observes and with which it interacts lie in fact on the same plane surface of the floor on which the robot is moving. For the same reason, the transform can be effectively used also in traditional fixedcamera surveillance systems. 3. Multi-sensor Systems In considering vision systems which rely on the use of more than one sensor, a first distinction between two broad categories of systems can be made. The first category comprises those systems in which the presence of many sensors simply aims at achieving an extension of the field of view or, more in general, an extension of the quantity of information which can be acquired: information acquired by each sensor is usually (pre-)processed independently, and only at a later stage are results of independent processing jointly evaluated. The second category, on the contrary, involves those systems, such as stereo systems, in which data acquired by one sensor is typically relevant to the application only if processed concurrently with data coming from the other sensors. In the following sections we will provide examples of multi-sensor

12

G. Adomi, S. Cagnoni, and M. Mordonini

systems belonging to the above-mentioned categories, with particular regard to the vision system of a robot goal-keeper designed to compete in the RoboCup middle-size league, and to a hybrid vision system with stereo capabilities which is being tested for both robotics and surveillance applications. 3.1. Conventional Multi-camera Systems In several applications, in which accurate and fast action is required from a mobile robot, there is the need for vision systems which provide, at the same time, wide field of view, high resolution and limited distortion. In practical terms, this means offering the opportunity to focus onto the objects of .interests instantaneously, to analyse them in details, and to be able to dedicate as much computation power as possible to image processing and understanding algorithm by limiting the cost of camera control strategies. Such goals can be achieved by multi-camera system, based on conventional sensors, in which the field of view of each of the N sensors covers about l/N* of the scene of interest. This is the case, for example, of the vision sensor of Galavron, the robot goal-keeper [4] we designed to compete in the RoboCup middle-size league. Figure 5 shows the goal-keeper. The robot vision system is based on two wide-angle cameras, that allow the visual field to be more than 180° wide. The cameras are placed on top of the goalie. The fields of view of the two cameras, each of which extends by about 70° vertically and by about 110° horizontally, overlap in a region about 20° wide which corresponds to the centre of the zone that lies immediately in front of the robot. As anticipated, to simplify camera control and co-ordination as much as possible, the two cameras are connected to two frame grabbers, which can be operated independently of each other. It is crucial for goal-keeper operation that no frame in which the ball is visible be lost.

Figure 5. The robot goal-keeper Galavrdn and the field of view of its vision system.

Notes and Reflections on Vision Systems for Mobile Robotics

13

An extremely simple acquisition strategy can be implemented as follows: when the ball cannot be detected by either of the two cameras, the camera from which the image is acquired is switched at each new frame acquisition. When the ball is detected by one of the two cameras the image is acquired from that camera until it enters the overlap region. At that point the other camera is switched on for acquisition. Such a strategy can be easily extended to systems having more than two sensors. Absolute localization and repositioning are based on vision, too. Each time the ball is not visible the goalie enters the repositioning mode, and uses the front line of the goal-keeper area to determine its location and to reposition itself in the middle of the area just in front of the goal. To do so, an efficient algorithm is used, based on a comparison between the orientation and intercept point of the lines delimiting the goal-keeper area, as acquired by the left and the right camera, as shown in Figure 6. To detect lines, the first step which is performed is colour segmentation of the image, followed by edge detection. Only the edges of the white areas are then transformed into the Hough domain [11], to allow for an easy computation of the parameters of the line equation. Once the parameters are known and the intersection between the visible lines is detected, the following parameters: 1. coordinates of goal-keeper area vertices; 2. position of goal relative to the robot; 3. distance from goal are used to detect one of 24 possible positions, which have been previously identified, for each of which a different closed-form solution to the localization problem, computed on such parameters, exists. 3.2. Conventional Stereo Systems Stereo vision is usually obtained by two cameras slightly displaced from each other, thus being characterised by having a widely overlapping field of view, or through the acquisition of two images from a single sensor that can move to simulate the availability of two cameras displaced as above. The sensors can be traditional cameras or even omnidirectional sensors [10]. The displacement of the two cameras generates two images of the same scene, taken from two different points of view. By comparing the two images it is therefore possible to compute stereo disparity and, from such data and from knowledge of the geometry of the acquisition system, to infer some threedimensional properties of the objects in the scene.

14

G. Adorni, S. Cagnoni, and M.

Mordonini

Figure 6. Above, from top: image acquisition, color segmentation, line detection. Below: one of the 24 possible solutions to the localization problem.

3.2.1.

Computing stereo disparity

Figure 7 shows a general case study for stereo systems: the same point P is observed from two points of view 0 1 and 02. Stereo disparity D can be defined as the distance between the two projections of the obstacle cast from 0 1 and 0 2 on the ground plane (the z=0 plane), i.e., the two points of coordinates (xpi,ypi50) and (xP2,yP250). Moreover, it is interesting to represent the disparity D as

Notes and Reflections on Vision Systems for Mobile Robotics

15

composed by the two components Dx and Dy. For a binocular sensor with two conventional cameras, Ol and 02 are represented by the optical centres of the cameras and the stereo disparity is independent of the orientation of the cameras. Setting the position of the two cameras, it is possible to compute Dx and D y as functions of the coordinates of P: Dx = h - ( | z 1 - z 2 | / | z 1 z 2 | ) - | X - X 0 | = h - K - | X - X 0 | Dy = h - ( | z 1 - z 2 | / | z 1 z 2 | ) - | Y - Y 0 | = h - K - | Y - Y 0 | where (xi,yi,zi) and (x2,y2,z2) represent the coordinates of 01 and 02, respectively, (X,Y,h) the coordinates of P, X0 =(x2Zi - xiz2)/(zi - z2) and K is a constant. This result shows that there is a straight line (of equation X = Xo) along which any obstacle produces a null disparity independently of its height. A similar result can be obtained for Dy. 3.2.2. IPT-basedobstacle detection The idea of using inverse perspective for obstacle detection was first introduced in [17]. If one applies the inverse perspective transform (IPT) with respect to the same plane to a pair of stereo images, everything that lies on that plane looks the same in both views, while everything that does not is distorted differently, depending on the geometry of the two cameras through which the stereo pair is acquired, This property is particularly useful for tasks in which a relevant

Figure 7. Schematic representation of the stereo disparity D obtained by looking at the point P from the two points of view Ol and 02. The x-y axis plane is the ground reference plane.

16

G. Adomi, S. Cagnoni, and M.

Mordonini

reference plane can be easily found. This is the case for navigation, either for vehicles travelling on roads (see [6,7,18] for a review on the subject) or for indoor-operating autonomous robots [10]. Three steps are therefore usually required to detect obstacles based on stereo vision: • application of the IPT to each of the two images; • subtraction of one image from the other one to compute differences; • remapping of the regions where obstacle candidates can be found on at least one of the acquired images, to label pixels either as obstacle or free space. 3 3 . Omnidirectional Stereo Systems The same considerations made regarding stereo systems made up of conventional sensors can be extended to systems in which the two sensors are catadioptric omnidirectional sensors. The only difference is the fact that, instead of the optical centre of the cameras, in this case we will have to refer to the point on the mirrors which correspond to the reflection of the point of interest. As will be shown more formally in the following, in this case, the positions of the two reference points 0 1 and 02 depend on the coordinates of P. Figure 8 shows the simulation of a stereo omnidirectional systems and of two virtual images of a RoboCup environment. The two mirrors of the simulated system are co-axial. In this case, the property by which rotation-symmetric mirrors with a discontinuity in the vertex can reduce reflection of lower objects which cross the axis up to making it collapse into a single point, is essential for this kind of system to work properly. Therefore, this constraint will have to be added to the specifications, when designing such systems.

Figure 8. The simulation of a possible omnidirectional stereo systems with two views of a virtual RoboCup scene.

Notes and Reflections on Vision Systems for Mobile Robotics

17

3.4. Hybrid Omnidirectional/pin-hole Systems (HOPS) HOPS [3] (of which two prototypes are shown in Figure 9) is a hybrid vision sensor that integrates omnidirectional vision with traditional pin-hole vision, to overcome the limitations of the two approaches. If a certain height is needed by the traditional camera to achieve a reasonable field of view, the top of the omnidirectional sensor may provide a base for the conventional sensor that can lean on it or be set aside it, as shown in Figure 9. In the prototype shown to the right, the conventional camera looks down with a tilt angle of about 60° with respect to the ground plane and has a field of view of about 80°. To obtain both horizontal and vertical disparity between the two images, it is positioned off the centre of the device. The 'blind sector' caused by the upper camera cable on the lower sensor is placed at an angle of 180° with respect to a conventional 'front view', in order to relegate it to the back of the device. If a lower point of view is acceptable for the traditional camera, the camera can also be placed below the omnidirectional sensor, provided it is out of the field of view of the latter, as in the prototype on the left. An example of the images that can be acquired through the two sensors of the first prototype to the left of Figure 9 is provided in Figure 10.

Figure 9. Two prototype HOPS systems.

G. Adorni, S. Cagnoni, and M, Mordonini

18

Figure 10, Example of images that ean be acquired through the omnidirectional sensor (left) and through the conventional camera (right) of HOPS.

The aims with which HOPS was designed are accuracy, efficiency and versatility. The joint use of a conventional camera and of an omnidirectional sensor provides HOPS with different and complementary features: while the conventional camera can be used to acquire detailed information about a limited region of interest, as happens in human foveal vision, the omnidirectional sensor provides wide-range, but less detailed, information about the surroundings of the system, as in human peripheral vision. HOPS, therefore, suits several kinds of applications as, for example, selflocalization or obstacle detection, and makes it possible to implement peripheral/foveal active vision strategies: the wide-range sensor is used to acquire a rough representation of a large area around the system and to localize the objects or areas of interest, while the conventional camera is used to enhance the resolution with which these areas are then analysed. The different features of the two sensors can be exploited in both a stand-alone fashion as well as in a combined use. In particular, HOPS can be used as a stereo sensor to extract three-dimensional information about the scene that is being observed. 3.4.1.

Computing disparity in HOPS

With reference to Figure 7, in the case of HOPS, 01 is the optic center of the conventional camera, while 0 2 is the point on the mirror surface that reflects the image of point P onto the camera in the omnidirectional sensor. This means that it is possible to reason as in the case of two traditional cameras, with the only difference that now the position of one of the two reference points 01 and 0 2 depends on the coordinates of P. In this case we obtain that: Dx = h • ( | Zl - f(X,Y,h)| / | Zl • f(X,Y,h)| ) • |X - X0(X,Y,h)| D y = h • ( |z, - f(X,Y,h)| / |z, • f(X,Y,h)| ) • |Y - Y0(X,Y,h)|

Notes and Reflections on Vision Systems for Mobile Robotics where f(X,Y,h) is the z coordinate of 02, function of the position of P. The only difference with the result reported above is that there exists a different straight line of null disparity for every possible position of the obstacle. As just shown above, the position of this line depends only on the coordinates of 0 1 and 02. If the conventional camera had been placed on the plane (x=c) passing through the symmetry axis of the omni-directional sensor and parallel to the plane x=0, X^ would equal c where X equals c, so that all the central part of the visible area would have null or very low stereo disparity. For this reason it is advisable that the conventional camera be placed as off this plane as possible, when stereo disparity is expected to provide useful information. 3.4.2. 3.4.2. Obstacle detection using HOPS The generalization of IPT to any kind of sensor, as discussed in a previous section, makes it possible to use HOPS as a stereo sensor for IPT-based obstacle detection, as previously described. This opportunity was tested in our laboratory by installing one of HOPS prototypes on a mobile robot and using it as its only sensor for navigation. In the experiments we performed, the robot had to reach a goal moving on a floor on which a few objects were scattered. Some of these were just flat objects, like sheets of paper, which would not interfere with robot motion. Some other objects, instead, had significant height, and were to be avoided by the robot.

Teat 1

Tmt 2 Figure 11. Test 1: (left) a large obstacle and an apparent obstacle separate the robot from the goal; (centre) the robot avoids the large one and ignores the apparent one; (right) and reaches the goal. Test 2: (left) the robot avoids the notebook bag; (centre) and the waste-paper basket, ignoring the sheet of paper; (right) and reaches the goal.

19

20

G. Adorni, S. Cagnoni, and M. Mordonini

Two experiments were carried out. In the first one, the robot had just to circumvent a large obstacle, while, in the second one, the robot had to pass between two smaller objects. In both cases the robot was required to ignore all 'apparent obstacles'. In particular the robot was to ignore all flat objects lying on the floor. In fact, they appear and would be treated in the same way as actual obstacles if only two-dimensional information is taken into account, due to their color, which is different from the floor color. Figure 11 shows a few frames taken from the video of this experiment. 4. Final Remarks In this chapter we have tried to offer a panoramic view, mainly based on examples of practical applications, of the different kinds of vision systems which can be used in mobile robotics. We have mainly focused on less conventional systems, such as omnidirectional systems and multi-sensor systems, which offer stereo vision capabilities. Specific attention was devoted to a peculiar hybrid configuration, we have termed HOPS, in which a conventional camera is associated to an omnidirectional one. This configuration is aimed at developing applications in which the properties of both kinds of sensors are requested. Besides their use as complementary sensors, we have discussed the use of HOPS as a stereo sensor. Although, because of its practical-example based nature and to our specific interests, the chapter is far from offering a complete and homogeneous review of the field, it provides a very general introduction to vision sensors for mobile robotics, while showing some original results obtained with unconventional ensembles of different kinds of sensors. References 1. G. Adorni, S. Cagnoni, and M. Mordonini, Proc. Asian Conf. on Computer Vision, 601 (2000). 2. G. Adorni, L. Bolognini, S. Cagnoni, and M. Mordonini, Proc. 7th AI*IA Conf., Springer LNAI2175, 344 (2001). 3. G. Adorni, S. Cagnoni, M. Mordonini, and A. Sgorbissa, Proc. OMNIVIS03, IEEE (2003). (available only on CD-Rom or online at http://www.cs.wustl.edu/pless/omnivisFinal/cagnoni.pdf) 4. G. Adorni, S. Cagnoni, S. Enderle, G. Kraetschmar, M. Mordonini, M. Plagge, M. Ritter, S. Sablatnog, and A. Zell, J. Robotics and Autonomous Systems, 36 n. 2-3, 103 (2001).

The full video can be watched and downloaded at http://www.ce.unipr.it/people/cagnoni/Filmati/caretta.mpg

Notes and Reflections on Vision Systems for Mobile Robotics

21

5. S. Baker and S. K. Nayar, in R. Benosman and S. B. Kang (eds.), Panoramic Vision: Sensors, Theory and Applications, Springer-Verlag Monographs in Computer Science, 39 (2001). 6. M. Bertozzi, A. Broggi, and A. Fascioli, Image and Vision Computing Journal, 16 n.8, 585 (1998). 7. S. Bohrer, T. Zielke, and V. Freiburg, Proc. Intelligent Vehicles '95, 276 (1995). 8. A. Bonarini, P. Aliverti, and M. Lucioni, IEEE Trans, on Instrumentation and Measurement, 49 n.3, 509 (2000). 9. L. Delahoche, B. Marie, C. Pegard, and P. Vasseur, Proc. lEEE/RSJ Int. Conf. on Intelligent Robots and Systems, 718 (1997). 10. C. Drocourt, L. Delahoche, C. Pegard, and C. Cauchois, Proc. IEEE/RSJ Int. Conf. on Intelligent Robots and Systems, 960 (1999). 11. N. Guil, J. Villalba, and E.L. Zapata, IEEE Trans, on Image Processing, 4 n . l l , 1541 (1995). 12. J. S. Gutmann, T. Weigel, and B. Nebel, Proc. 1999 IEEE/RSJ Int. Conf. on Intelligent Robots and Systems, 1412 (1999). 13. R. A. Hicks and R. Bajcsy, Proc. 2nd Workshop on Perception for Mobile Agents, 82 (1999). 14. H. Ishiguro, in R. Benosman and S. B. Kang (eds.), Panoramic Vision: Sensors, Theory and Applications, Springer-Verlag Monographs in Computer Science, 23 (2001). 15. F£. Ishiguro, K. Kato, and M. Barth, in R. Benosman, and S. B. Kang (eds.), Panoramic Vision: Sensors, Theory and Applications, Springer-Verlag Monographs in Computer Science, 377 (2001). 16. P. Lima, A. Bonarini, C. Machado, F. Marchese, F. Ribeiro, and D. Sorrenti, J. of Robotics and Autonomous Systems, 36 n. 3, 87 (2001). 17. H. A. Mallot, H. H. Bulthoff, J. J. Little, and S. Bohrer, Biolog. Cybern., 64, 167 (2001). 18. K. Onoguchi, N. Takeda, and M. Watanabe, IEICE Trans. Inf. & Syst, E81-Dn.9, 1006(1998). 19. M. Plagge, R. Gunther, J. Ihlenburg, D. Jung, and A. Zell, RoboCup-99 Team Descriptions, 200 (1999) {available at http://www.ep.liu.Se/ea/cis/l 999/006/cover.html). 20. T. Sogo, H. Ishiguro, and M. Trivedi, in R. Benosman and S. B. Kang (eds.), Panoramic Vision: Sensors, Theory and Applications, SpringerVerlag Monographs in Computer Science, 359 (2001). 21. Y. Yagi, S. Kawato, and S. Tsuji, IEEE Trans, on Robotics and Automation, lOn.l, 11 (1994).

This page is intentionally left blank

23

A N O T E ON P I C T O R I A L I N F O R M A T I O N VITO D I G E S U Dipartimento

di Matematica ed Applicazioni, Universita di Palermo Via Archirafl 34, 90123, Palermo

The chapter introduces new entropy measures that use the image information content such as grey levels and their topological distribution in the image domain in order to perform the classification of the image itself. The main aim of the chapter is to study the role of the image entropy in perceptual tasks and to compare the proposed approach with others wellknown methods. Experiments have been carried out on medical images (mammograms) due to their variability and complexity. The image entropy approach seems to work quite well and it is less time-consuming if compared with the other methods.

1. Introduction Visual science is considered one of the most important fields of investigation in perception as a matter of fact that a large part of the empirical knowledge we have about the world is based on the visual perception. In order to fully understand vision mechanisms we need to consider the interaction and the integration among all sensors of perception (e.g. hearing, touch, sense of smell). Moreover, the information that comes from our sensors to the brain is elaborated by using also mental models (mental information) that are stored somewhere. Furthermore, it has been always a challenging problem to understand how we see and how we interpret visual scene surrounding us. One question arises What processes fill the gap between the picture-like information on the retina and conceptually structured perceptual representations! Visual perception is an interesting investigation topic because of its presence in most human activities. It can be used for communication, decoration and ritual purposes. Visual information depends on the context. Images may transfer emotion (see Figure 1). We could call this kind of information emotional.

Figure 1. Women Playing Music (Tintoretto 1518-1594, Venice).

V. Di Gesu

24

Figure 2. Graffiti representing acrobats (Addaura caves, Palermo, Italy).

On the other hands, images may increase our knowledge; for example, the scene of hunting men represented by graffiti on walls prehistoric caves in Figure 2 tells us about the style of life of prehistoric men. We call this kind of pictorial information knowledge acquisition. Note that, graffiti can be considered the first example of visual language that uses an iconic technique to pass on history. They may also suggest us how prehistoric men internalized external world. The intrinsic meaning of an image depends also on the expectation and the resulting perceived may influence our decision, this is the case of medical diagnostic (see Figure 3). In this case pictorial information allows us to discover a cue. On the other hand, a static scene may transfer dynamic information telling us how the world will evolve. This is the case of the cartoon photogram in Figure 4 where the lion is approaching its cub. It is proud and loving, we could image the discussion even without the help of sound.

Figure 3. CTA image of a patient's abdomen.

A Note on Pictorial Information

25

Figure 4. Dynamic information from a single photogram.

Verbal messages can be added to pictorial data that involve the integration of different sources of information (see Figure 5). All these examples do not exhaust the taxonomy related to pictorial information, but they highlight how complex is the definition of information when the semantic is included in the problem. In this case, the computation of a system information by means of the entropy as defined by Shannon [1] could be unsatisfactory. In fact, emotional or linguistic features can not be assimilated to those used in thermodynamics, even if they still characterize changes in the system status. Subjective experience may play a fundamental role whenever inexactness and vagueness can not be modeled by probabilistic reasoning; in fact, cloudy features may exist that are not describable in terms of probability distributions. For example, the beauty of some thing or the tallness of a man. In all these cases, education, fashion and global knowledge may play a crucial role in making decisions. The problem of modelling non numerical information is of great interest in artificial systems. Soft computing and fuzzy sets can provide a powerful mathematical tool to model abstract concepts, allowing the definition of functions that satisfy entropic properties.

W§ help ynu Figure 5. Linguistic and visual integration.

26

V. Di Gesit

Fuzzy Set Theory was formalised by Zadeh at the University of California in 1965 [2]. What Zadeh proposed is very much a paradigm shift that first gained acceptance in the Far East and its successful application has ensured its adoption around the world. Crisp concept of true and false are replaced by continuous degree of true ranging in the interval [0,1]. Quoting Zadeh: ... For a long time humankind has endeavoured to understand the laws of the surrounding world and has made continuous attempts to describe the phenomena occurring in the world. Naturally we want to achieve the most adequate descriptions by means of exact and precise terms. Mathematical language should be the best tool to express such descriptions; however, the language of set theory and extensional logic is sometimes insufficient...

The framework of fuzzy sets allowed us to define entropy based functions that have been used in image and signal analysis to perform deconvolution [3] and segmentation [4], to measure the pictorial information [5], and to define image differences [6,7,8]. Entropic measure has been used for medical diagnosis in [9,10]. 2. Processing Pictorial Information Pictorial information processing (PIP) is developed throughout several layers of increasing abstraction that corresponds to a set of iterated transformations. One of the main tasks of a PIP is the ability to focus the computation in areas of interest, based on the maximization of an expected utility criterion that includes costs and benefits. This feature is also named visual attention or selective vision. Visual attention is also included in natural vision system, and it allows to reduce the computation time, avoiding redundant computation [11]. Moreover, PIP should be able to adapt its behavior depending on the current goal and the nature of the input data. Such flexibility can be obtained in systems able to interact dynamically with the environment. The term active vision has been coined, to address such kind of visual computation [12]. In the active vision paradigm, the basic components of the visual system are visual behaviors tightly integrated with the actions they support. Because the cost of generating and updating a complete, detailed model of the environment is too high, the development of this approach to vision is vital for achieving robust, real-time perception of the real world. The search of regions of interest is usually followed by a matching procedure that performs objects classification or the clustering of relevant features. There are a number of feature dimensions (e.g., position, colour, size, orientation, gradient, form, trajectory, etc.) that govern the grouping of elements of the receptive field into objects by the following Gestalt rules: neighbouring

A Note on Pictorial

27

Information

elements with similar features tend to belong to one and the same object; proximity and similarity are the main factors in perception [13]. It follows that the extraction of information depending on grouping rules may be defined on the basis of the problem we are considering. In the case of visual information graphs can be associated to each level of the recognition process. Moreover, visual pattern recognition is a process that is performed throughout several layers of increasing abstraction, corresponding to a set of iterated transformations. Therefore, models of visual processes are developed through a hierarchy of graphs each corresponding to a different feature space. For example, Di Gesu and Zahn in [14] developed a new method to recognize lines based on hierarchy of different space starting from a set of points in two-dimensional space. Here, curves are considered as ordered sequences of lines or edges as primitive elements of a recognition system. The algorithm firstly detects short sequences of closely spaced points by using single link algorithm based on the Minimum Spanning Tree. The result is a set of very oblong (nearly linear) shapes, named segments. Then, segments are linked together into longer curves with slowly varying directionality. The clustering of segments is performed by a KNN link algorithm that uses a measure of closeness or similarity between segments based on their reciprocal orientation and spatial position. This kind of procedure can be extended by considering hierarchies of space of increasing complexity (see Figure 6).

Ptfarti

% >

Lines

Curve* Figure 6. Hierarchical computation from points to curves.

28

V. Di Gesii

2.1. Soft Vision Fuzzy sets provide methods for the analysis of images [15]. Their application cover most of the analysis, from low to high level vision. For example, fuzzy-convolvers have been designed whose kernel values depends on both the point spread function of the detector and the local pixelrelations. Extension of mathematical morphology to grey levels images has been carried out by using fuzzy-sets in [16]. This extension is based on the interpretation of the gray level of a given pixel as its belonging degree to X, and by using min/max operators. High level vision process could be considered as a complex pattern matching in the sense that recognition is the evaluation of a fitting algorithm applied to heterogeneous information (visual and textual). This fitting is realized throughout the visit of pictorial data-base and knowledge-base. Moreover recognition systems must be able to handle uncertainty, and to include subjective interpretation. For example, an intelligent system should be able to evaluate and understand the following fuzzy-propositions: a) the chair, beyond the table, is small; b) the chair, beyond the table, is very small; c) the chair, beyond the table, is quite small; d)few objects have straight medial axis. Where small is a fuzzy verbal predicate, very and quite are fuzzy attributes. The evaluation depends on the meaning that is given to small, very and quite. Moreover the objects chair and table, and the relation beyond must be recognized with some degree of truth. The proposition d) contains the fuzzy quantifieryew. 3. Applications 3.1. Image Segmentation The main problem in image analysis is to extract the signal from the background and the selection of the important compounds. This phase is also said image segmentation. The success of this phase may influence all recognition process. The dog in Figure 7 shows how the segmentation process is complex and not easy to implement in artificial systems.

A Note on Pictorial

29

Information

Figure 7. The hidden dog.

Pre-attentive processes may help in performing image segmentation. For example the applications of local symmetry operators may help to extract proper segments from very complex images. The reason is that attentive processes are able to focus the attention on the relevant regions of the scene reducing the redundancy of the image data. For example, in Figure 8a the operator DST (Digital Symmetry Transform) [17] is applied to the image in Figure 1. Figure 8b shows the regions of interest that have been highlighted by the DST. It is interesting to note that the entropy of the input image, X, of dimensions A/xMis usually greater than the entropy of the pre-processed DST(X), where the entropy used is the classical one, with the gray levels, pi}, interpreted as a measure of probability of the pixel (i,j):

(a)

N

M

i=\

j=\

(b)

Figure 8. Pre-attentive process: (a) DST of the painting in Figure 1; (b) selected regions of interest.

30

V. Di Gesu

In our example H(X)=524 and H(DST(X))=2539 note that the entropy of the image, S, containing the selected regions is H(S (A!))=2.63. This means that redundancy has been reduced maintaining a good image quality in the interesting parts. 3.2. Texture Analysis In the previous example the entropy is computed in the whole image. In some cases entropy functions can be computed locally in sub-images, A, of X. This kind of computation allow us to detect local structure of the image and the application of the local entropy operator is able to detect non regular textures. For example, in Figures 9(a,b) are shown the input image, X9 and its DST. Figures 9(c,d) show the effect of cutting the gray levels less than a threshold q»0. Regular textures can be detected using a measure of entropy that is function of a given correlation step x (in pixels). We call this information measure correlated entropy.

H{x) = Y,f{x)®f{y) XEX

(a)

y = (gy, ix + T>J.x + T)

(b)

(d) Figure 9. (a)*;

0.3

0.4

4...

0.6

L

1

t

OS

07

0.8

OS

Figure 11. Trend of the entropic similarities with 5.

A Note on Pictorial Information

33

Table 1. Comparison of Go with classical measure of similarities.

Similarity Go D CO AD

Correct matches 92.6 81.5 80.0 75.5

Mismatches 7.4 18.5 20.0 24.5

4. Discussion and Final Remarks The chapter analyzes the concept of visual information from several perspectives. Information is associated to changes in the status of a system. The definition of the information content of a system in terms of its entropy computed in the phase space is not always satisfying. For example, emotional, and linguistic information can't be computed via entropy functions. In these cases fuzzy-sets could be a useful tool to represents abstract thinking in a computer. On the other hand, the definition of entropic functions based on the properties of convex and not decreasing functions are appropriate for handling numerical properties of an image. The problem of analyzing pictorial information in this cases can be formulated as the alternation of attentive processes and matching algorithms. For example, several low-level visual tasks can be handled using measures of local information. Entropic distances are useful in image matching problems; they show an interesting property of expansion/compression, that can be usefully used to group in finer classes elements close to the origin and to group in broadest classes elements close to the frontiers of the distance space. References 1. C. E. Shannon, "A mathematical theory of communication," Bell System Technical Journal, 27, pp. 379-423 and 623-656, July and October, (1948). 2. L. A. Zadeh, "Fuzzy sets", Inform. Control, 8, pp.338-353, (1965). 3. F. Kossentini, M.J.T. Smith, C.F. Barnes, "Image coding using entropyconstrained residual vector quantization", in IEEE Transactions on Image Processing , 4 , pp.1349-1357, (1995). 4. Z. Xiong, K. Ramchandran, and M. T. Orchard, "Space-frequency quantization for wavelet image coding", in IEEE Transactions on Image Processing, 6, N.5, pp. 677-693, (1997). 5. J. Skilling, R.K. Bryan, "Maximum entropy image reconstruction: general algorithm", Month. Notices Roy. Astronom. Soc, 211, pp. 111-124, (1984). 6. V. Di Gesu, S. Roy, "Pictorial indexes and soft image distances", in Lecture Notes in Computer Science, N.R. Pal and M. Sugeno (Eds.), pp. 200-215 (2002).

34

V. Di Gesii

7. V. Di Gesu, S. Roy, "Fuzzy measures for image distance", in proc of Advances in Fuzzy Systems and Intelligent Technologies, F. Masulli, R. Parenti, G. Pasi (Eds.), Shaker Publishing, pp.156 -164, (2000). 8. J. Zachar, S.S. Iyengar, "Informaton theoretic similarity measures for content based image retrieval", Journal of the American Society for Information Science and Technology, 52, N.10, pp. 856-857, (2001). 9. A. Casanova, V. Savona and S. Vitulano, "Entropy As A Feature In The Analysis And Classification Of Signals", MDIC, (2004). 10. A. Casanova, V. Di Gesu, G. Lo Bosco, S. Vitulano, "Entropy measures in Image Classification", Human and Machine Perception 4: Communication, Interaction, and Integration, V. Cantoni et al. (Eds.), World Scientific, Singapore, (2005). 11. CM. Brown: Issue in selective perception. In Proc. 11th IAPR Int. Conf. on Patt. Recog., IEEE Computer Society Press, Vol. A, pp.21-30, (1992). 12. Promising direction in active vision. Tech. Report CS 91-27, M.J. Swain and M. Strieker (Eds.), University of Chicago, (1991). 13. M. Wertheimer, "Untersuchungen zur Lehre von der Gestalt I. Prinzipielle Bemerkungen ", Psychologische Forschung, 1, pp. 47-58, (1922). 14. V. Di Gesu, C.T. Zahn, "A general method to recognize two-dimensional dotted curves", in Radiol. Clin.No.Amer, S.L.A.C Tech.Rep. 75/01, Stanford University, (1975). 15. S.K. Pal and D.K.D. Majumder, "Fuzzy mathematical approach to pattern recognition", Jon Wiley & Sons, (1986). 16. V. Di Gesu, "Artificial Vision and Soft Computing", in Fundamenta Informaticae, (37), pp.101-119, (1999). 17. V. Di Gesu, C. Valenti, "Symmetry operators in computer vision", in Vistas in Astronomy, Pergamon, 40(4), pp. 461-468, (1996). 18. B. He, I. Ounis, "Inferring Query Performance Using Pre-retrieval Predictors", in 11th Symposium on String Processing and Information Retrieval (SPIRE 2004), October 5-8, Padova, Italy, (2004). 19. A. Del Bimbo, E. Vicario, S. Berretti, "Spatial arrangement of color in retrieval by visual similarity", in Pattern Recognition, 35(8), pp. 1661-1674 (2002). 20. P.M. Kelly, T.M. Cannon, and D.R. Hush, "Query by image example: the CANDID approach", in Proc. of the SPIE: Storage and Retrieval for Image and Video Databases III, 2420, pp. 238-248, (1995). 21. E. Ardizzone, V. Di Gesu, M. La Cascia, C. Valenti, "Content Based Indexing of Image and Video Databases", in Proc. 12th ICPR, IEEE, (1996). 22. M. La Cascia, E. Ardizzone, "JACOB: Just a content-based query system for video databases", Proc. oflCASSP '96, Atlanta, (1996). 23. V. Di Gesu, V. Starovoitov, "Distance-based functions for image comparison", in Pattern Recognition Letters, 20, pp. 207-214, (1999).

This page is intentionally left blank

35

BIOMEDICAL SIGNAL PROCESSING AND MODELING: MULTISCALE AND MULTIORGAN INTEGRATION OF INFORMATION SERGIO C E R U T T I 1 Department of Bioengineering

Polytechnic University, Piazza Leonardo da Vinci 32, 20133 Milano, Italy

Biomedical signals cany important information about the behavior of the living systems under studying. A proper processing of these signals allows in many instances to obtain useful physiological and clinical information. Many advanced algorithms of signal and image processing have recently been introduced in such an advanced area of research and therefore important selective information is obtainable even in presence of strong sources of noise or low signal/noise ratio. Traditional stationary signal analysis together with innovative methods of investigation of dynamical properties of biological systems and signals in second-order or in higher-order approaches (i.e., in time-frequency, timevariant and time-scale analysis, as well as in non linear dynamics analysis) provide a wide variety of even complex processing tools for information enhancement procedures. Another important innovative aspect is also remarked: the integration between signal processing and modeling of the relevant biological systems is capable to directly attribute patho-physiological meaning to the parameters obtained from the processing and viceversa the modeling fitting could certainly be improved by taking into account the results from signal processing procedure. Such an integration process could comprehend parameters and observations detected at different scales, at different organs and with different modalities. This approach is reputed promising for obtaining an olistic view of the patient rather than an atomistic one which considers the whole as a simple sum of the single component parts.

1. Introduction The processing of biomedical signals is an important step towards the obtaining of objective data from the analysis of living systems. The aims could be: i) to improve the physiological knowledge of the system; ii) to provide quantitative data for clinical purposes (diagnosis, therapy and rehabilitation) [7] [17] [24] [25]. A wide variety of processing algorithms are traditionally applied to biomedical signals. Due to the fact that such signals are often quite difficult to be processed and characterized by low signal/noise ratio, practically all the methods introduced in the signal processing arena have been more or less applied in the different approaches of biomedical signal processing: stationary and non stationary signal analysis, second-order or higher order approaches,

Work partially supported by a EU Grant, My-Heart Project, 2004.

36

S. Cerutti

with deterministic or stochastic methods, in time domain and in frequency domain, using linear and non linear algorithms, etc. New paradigms related to a new concept of biomedical signal processing are considered important. In the following, an approach is described which considers signal processing together with modeling of the relevant biological systems under studying, as well as an integration of the real information which is obtained at different scales, using different investigation modalities, with different organs. An integrated view of all this obtained information, generally realized through a multidisciplinary pathway which involves various and different competences and professionals, might produce a precious enhancement of information, thus providing a better knowledge of the underlying systems. 2. Signal Processing and Modeling of Biological Systems A basic issue of innovative biomedical signal processing procedures is that a stronger link is required between modeling of biological systems and processing of the signals involved [3][5][9]. Most often, in fact, researchers who do modeling do not do signal processing. A well-trained biomedical engineer is capable to integrate the knowledge which is required to deal with models and with signal processing procedures and algorithms. A major innovative tool would be the improvement of medical knowledge through a cooperative process between signal processing and modeling, along the way indicated in Figure 1. Noise & disiuibancies

Biological process

MeanueniLsyst: detection, measur. and pre-processing

Processing

Signal+noi5ft

i

i

B Model

Processing parameters

Signal^ Data from hfi model Model structure #1

I

Comparison



Medical knowledge production

Processing Model parameters

Model structured

Figure 1. Integration between biomedical signal processing (channel A) and biological modeling (channel B) to create medical knowledge.

Biomedical Signal Processing and Modeling

37

We may define one channel A to produce information (upper part of the figure) which is connected to the processing of biological signals and one channel B (lower part of the figure) which is connected to the modeling of the relevant biological system. The "processing" block takes information from the signals derived from the biological process under studying, as well as from the modeling of the physiological system. As an example, we may consider cardiovascular system (and in particular the physiological mechanisms which make the heart rate to change on a beat-to-beat basis). The signals involved may be heart rate variability (HRV) and/or arterial blood pressure variability (ABPV), while we have many physiological models in literature which describe the HRV and/or ABPV. Obviously, the signals have to be properly chosen and processed as well as the models suitably validated: at the end of the procedure a joint analysis of signal processing and modeling may indeed produce new information and knowledge, as depicted from the block in the right part of the figure. Another example of this important integration of processing and modeling is given by the next figures (Fig. 2, 3 and 4). Figure 2 shows a well known model by Koepchen (1986) [13] in which many complex physiological relationships are schematically introduced for the description of the control of autonomic nervous system on cardiovascular functions: this is considered a typical (complex) model which is designed by taking into consideration only physiological observations. Figure 3 shows instead how in [2] we attempted to identify where various rhythms which are present on cardiovascular signals could be generated at central, autonomic and vascular levels. LF (low frequency) and HF (high frequency) are two basic rhythms which are found on cardiovascular variability signals (variability in heart rate: HRV and variability in arterial blood pressure: ABPV) [23]. Further, three signals were selected as information carriers of the complex physiological mechanisms underneath: heart period variability (RR series from the ECG signal), systolic blood pressure variability (series of systolic values from arterial blood pressure signal) and respiration. The model by Koepchen and the evidenced biosignals made us to design a black-box model, indicated in Figure 4, which describes the interactions between HRV (signal t), ABPV (signal s) and respiration (signal r) in the autonomic control of cardiovascular function. Of course, the model takes into account only a small part of the information which is suggested by the physiological model where we started from; in any case, after a validation of the black model as done in [3], it is possible to provide important physiological meaning to the blocks reported in Figure 4 (estimation of respiratory sinus arrhythmia, calculation of baroceptive gain in a closed-loop way, measurement of mechanical effect of heart rate on arterial blood pressure, gain and phase of the feedback regulating mechanisms, etc).

S. Cerutti

38

AR AS MXtlttrois

am tmmsw

1

•UKUIM ttftout ItfflUS L _

8"

Xlltllill

ItWMRl Mvftfm

uststu «cs:in •HI

• (All ttKt

o^oh snui i i i r

o o

itwtEB wmn

Uui mriMi

IPS

• t i n s »ci»t«

iiiiiui lUMmriiis

— ctiiut t u r n

I HSKtW ' IHlllitMBlHf

SMWI« HfSUfS

Figure 2. A physiological model of the complex central and peripheral interactions of autonomic regulations. From [13].

an'.'.TC'.im!i'wiiiJiimw»

Figure 3. A model inspired to the structure of the physiological model reported in Figure 2. Here the signals from which to detect the information on autonomic regulations are evidenced (respiration, heart rate and arterial blood pressure). From [2].

39

Biomedical Signal Processing and Modeling

v% M,

w.

M,

,, vasomotor « "» modulation

respiration f 1

mechanical

H* Rt

HP

SAP

H,is sinus nodsML modulation

barorecepev9 mechanisms

H SS AP regulation

Figure 4. Block diagram of the black-box model describing the interactions between heart rate, arterial blood pressure and respiration in a double feedback loop, starting from the physiological considerations contained in Figs 2 and 3. From [3].

Another interesting approach of joint analysis of signal and model parameters is shown in Figure 5. Fig. 5a indicates a model originally developed by Kitney RI (1979) [12] and successively modified by Signorini (1997) [22], where there is a very simple description of the closed-loop regulation of blood pressure (signal y) according to different values of respiration signal (r). For simplicity r is assumed as a pure sinusoid indicated as equal to e sin(27tft). There is a negative feedback loop which induces oscillation in y signal due to the fact that there is also a non linear block in the forward path. According to the different values of the closed loop gain and the frequency of the respiration signal (by simplicity assumed as a pure sinusoid) we may have either a synchronization between respiration rate and blood pressure rate (i.e. they take the same frequency) or more complex behavior due to the non linear dynamics. In Fig. 5b there is the diagram of gain HI as a function of f (frequency of respiratory signals). It is possible to see that a very complex phenomenon is evidenced: there is an area of 1:1 locking (synchronization between respiration and blood pressure) in the lower part of the figure, but also various transitions (bifurcations), when increasing HI, from synchronization to other forms of locking (i.e. 1:3), or to torus, or to torus breakdown with also a chaotic behavior just in the normal frequency range of respiration (0.20-0.30 Hz). Processing of arterial blood pressure signals confirm that, in this range, arterial blood pressure may present a chaotic behavior. This paradigmatic statement of "chaotic" behavior in the autonomic control of

40

S. Cerutti

cardiovascular functions supports in many instances the hypothesis that the high complexity in the physiology of this system may be explained by a non linear dynamics. A non linear approach may thus be employed to better describe the model and hence a proper quantification of its behavior could be done through parameters of measurement of the strange attractor properties like Lyapunov exponents, Poincare plots, fractal dimension and so on [10][21]. Disclosure from chaotic behavior or decreasing of their characteristic parameters are generally correlated to pathological conditions [14][26][15]. This is certainly another significant example of how integration between models and signal processing are able to produced new important physiological information. Ml

+

f

)

Non linearity X

"1

/ m

\f

>

M(l+Ts)e (1+Ts)(l+T2s)

y

—•

(a) H,

(b) 1:1 phase lockup

I Neimark-Sacker |

0.10 0.15 / 0.20 0.21 0.30 f frequency of the respiration signal (HF)

/ Cycle bifurcation

0.40

Period one to period three

Figure 5. Simplified model of arterial blood pressure regulation (a), from [12] and modified by Signorini et al. [22], Diagram of Hi gain on the previous model as a function of the frequency f of respiration signal (b): a wide variety of complex behavior is noticed due to the non linear characteristics of the model, depending upon the value of f: phenomena of entrainment, bifurcation, synchronization and chaos are evidenced. For details, see text.

Biomedical Signal Processmg and Modeling

41

3. Multiorgan and Multiscale Integration Another important issue which is connected to biomedical signal processing and the modeling of the relevant biological systems involved, consists in the integration of different information sources which may indeed constitute a unique data base relative to that single patient. That information integration is able to put together data which derive: 1) from different biological system or organs (multiorgan), i.e. integrating cardiovascular with respiratory and with endocrynometabolic systems; 2) from different signal modalities (multimodality), i.e. making a "registration" of 3-D MRI cerebral data with 3-D EEG mapping, thus combining the high spatial resolution of the first technique with the high temporal resolution of the second one; 3) from different scale of information (multiscale), i.e. various pathologies may be better investigated by considering information at the various scales on which the phenomenon may be studied. Some examples of this important integration paradigm will be illustrated below: here it is fundamental to state that in many instances there is also the need of integrating different professional figures, cultural backgrounds, training processes, personal attitudes and so on. Therefore is not simply a rather "technical" data integration which may allow the fulfillment of significant innovative results, but a more complex process of integration of knowledge from various and synergetic perspectives. From one side, encouraging a cooperative research among physiologists, clinicians, biologists, bioengineers, biophysicists, biomathematicians, etc, but also among experts of signal processing with colleagues experts in medical images, in molecular biology, in modeling at the various possible scales: from DNA/RNA scale, to gene/protein, to cell and cell fibers, to organ, to multiorgan analysis, to the single patient intended as a whole and unique system. 3.1. Multiorgan Integration As an example of multiorgan integration, it is possible to cite the case of sleep studies. There, biological signal recordings, generally carried out during polysomnographic procedures, put into emphasis the fact that there is an involvement of several organs or systems. Figure 6 shows a case of a patient who manifests the so-called restless-leg syndrome (RLLS), i.e. he undergoes a period of sleep (in stage 2) in which he presents a movement of legs, almost periodically. What is more interesting is that during this period there is a remarkable synchronization among biological rhythms of different origin: passing top-down in the figure, the first panel shows EEG tracing, where there are well visible phenomena of arousals, with a strong synchronization with the bursts of EMG signal (second panel) recorded at the

S. Cerutti

42

level of tibial muscle (these bursts are obviously primarily associated with the RLLS). Further, the third panel shows the RR series (the beat-to-beat heart cycle) as detected from a standard ECG lead and the last panel shows the respiration signal, recorded via plethysmographic technique. Horizontal axis is the number of consecutive heart beats and all the four signals are aligned along this temporal occurrence. In order to obtain this common horizontal axis reference, the values of EEG and EMG in Fig. 6 are obtained as the mean values of instantaneous EEG and EMG taken between two consecutive cardiac cycle (RR's), while R is the respirogram, i.e. the value of respiration in correspondence with the R peak occurrence in ECG signal. A very important synchronization among the four signals is noticed which is almost intriguing in its physiological interpretation: EMG bursts seem strictly correlated with acceleration/decelerations in heart rate and also with a kind of modulation mechanism in respiration which is even related to arousal phenomena in central nervous system (EEG signal). Important possible causal relationships among the signals coming from different physiological compartments (central and autonomic nervous systems, cardiorespiratory system and muscular system) are worth to be deeply and in more detail investigated [4].

•- :#44«M»»»j « M MM»«tNiNi IHIK4 k^4m iHi'

I

ion

TOD

M O 4, Figure 2 (b) and (c). At last, supposing the test for domain D5 fails because during the update of the centroid position for c\ there will be domains in the corresponding cluster with distance greater than e, that represents the maximum radius for a cluster. In this case a new cluster is created with centroid c2 located at the coordinates of its domain D5, Figure 2 (d). In the same way if any centroid is found for a domain, in scanning the list of centroids, a new cluster is created. At last, in order to deal with small shifts of the eyes, nose and mouth in the extracted face region due to

166

A. F. Abate, M. Nappi, D. Riccio, and M. Tucci

the errors committed during the feature localization process, also a limited number of nearest neighbor of the original ranges corresponding to the entry points are considered. Starting from the current entry point all centroids are computed by means of the aforesaid algorithm, then with a spiral visit centered in the current entry point height neighbor of the range are considered. These neighbors are kept on an Archimede's spiral p = a-6 where p is the distance from the center to the tracing point, 0 is the angular distance covered and a is a fixed constant. With the spiral visit, higher the distance from the entry point minor the considered number of neighbors. This makes sense if it is taken in account that more far the neighbor from the current entry point, minor their similarity with the current range. Then avoiding the problem of small shifts the probability that the correct entry point falls near to the prefixed position is higher. Once the list of centroids has been computed, it has to be rearranged so that a distance function can be defined for the next comparisons. It is sensible that, comparing two face images the approximation error of a centroid belonging to former with the error of the nearest centroid in the latter is matched. However this makes the comparison task much expensive. Indeed each centroid in the list of centroids consists of two coordinates C(x,y), the nearness of two centroids can be esteemed by the euclidean norm. Let be L\ a list of centroids with length n=\Li\ to be compared with another L2, with length m=\L2\. For each centroid CkL] in L\ the nearest centroid C J L in L2 is searched and the absolute difference between the correspondent approximation errors | ekcL - e*cL | is computed. The computational cost of this operation is so 0(n-m). Of course it can be done better, if the centroids in the list L\ are organized in a KD-Tree spending 0(«log(«)), in log(w) time the searched centroid 0L. is retrieved, then the overall computational cost reduces to 0(nlog(n)+nlog(m)) = 0(n\og(n-m)). Now it can be tried to obtain a better result representing spatial location of centroids with the Peano keys. Given a centroid C(x,y), the correspondent Peano key ac is computed interleaving bits of x andy, from the less significant digit to the most one. From the literature it is known that Peano keys are useful in indexing tasks because mapping a 2D space in a ID space, they preserve most of the spatial information of the original data. In the next step the computed keys are sorted, this can be done in a linear time Oin) with the Radix Sort algorithm. Making comparison between L\ and L2, 0(m) time is spent in order to search for the centroid CJi in L2 nearest to the first centroid CV in L\, holding memory of/V Then it is observed that, generally, the location of the next centroid in L2 nearest to a centroid CkL falls not so far from the position j \ , about j\+c, where

167

Recognizing Occluded Face Using Fractals

experimentally c has been found as about 0 < c < 10 and7'k is the location in L2 of the nearest of Ck_1L in Lu with k>\. Then for each centroids in Z,! it has to be tested only c centroids in L2, so that the overall complexity of the comparisons is now 0(n+c-m) = 0(n+m) that is linear because c is a constant. Having a low computational complexity making comparisons is crucial considering that in a huge database of face images, millions of images could have to be tested. Now the Peano curve rearrangement of the centroids and its consequent advantages are shortly examined. Let be L a list of centroids. Each centroid in L hold two main informations, that are the centroid coordinates C(x,y) and the approximation error ec of the domains in the correspondent cluster. ROF generates the Peano key for each centroid in L, interleaving the bits of x and y. Sorting these keys in a new list L according to the Peano curve there are two main advantages. In the first instance the distance between two feature vectors can be computed efficiently as described above. Second the information is organized in a one dimensional array, localizing eyes, nose and mouth in a fixed location. This is useful when a subject has to be authenticate only by means a subset of its facial features. Furthermore in this way, partial occlusion in the face image can be easily located. 3. Definition of the Distance Function A(A, B) In this section the distance function used comparing two different face features vector is defined. The dominion of this function consists of bi-dimensional vectors S eR2 where [a,b^eS with a a Peano key and b a real value representing the mean value of the approximation error for the centroid centered in a = P(x,y). Given two vectors S,T eR2 the operator iy(S,T) is defined as follows: ¥i{S,T)

bT

-b's

(2)

with / / ( £ , r ) = min a/ ~a's\ that is

/J(S,T)

represents the index in T of the point a'T = Pi{xl,yl)

the point a's = P2 [x2,y2)

in S .

nearest to

168

A. F. Abate, M. Nappi, D. Riccio, and M. Tucci

Then for each item as = Ps (x, y) eS

it is search for the nearest item

aT = PT (x,y) eT according to the euclidean norm JPS -PT|

and the quantity

\bs -bT\ is computed, that is the absolute difference of the approximation errors corresponding to the nearest points Ps e S and PT e T . At last summation on / of the values of y/t (S,T) is performed. In order to render the distance function robust with respect to the partial occlusions, it is observed that if yri{S,T') is too much great, it does not supply useful information, cause of a possible occlusion. Then supposing useful only the informations provided by values of i/fj(S,T) ranging from 0 and 2-m where m represents the average 1 |S| w = T—r-/^^(5,r). A threshold operation that cuts only the values greater than 2-m is applied, leaving unchanged those smaller. That can be done with the following function: (3)

\s\ '=1

\T\ '=I

where

. 7S

~S =

(s(i)-2-E[s])-\s{i)-2-E[S}\ 2-(s(i)-2-E[S]) {(ai,bi)GS3ris^0)

4. Experimental Results There are several standard databases used by the scientific community in order to assess the performances of the proposed algorithms in the field of the face authentication. Two of the most used face databases are FERET [4] and AR Faces [3], while recognition rate and CMS (Cumulative Match Score) have been used as a measure of the performances (the CMS measure has been defined in details in the FERET protocol [4]). There are two main aspects to be investigated evaluating performances in case of occluded faces. The first one is to estimate the contribution of each face element to the global recognition rate.

Recognizing Occluded Face Using Fractals

169

It make sense that eyes, nose and mouth have different weights in the recognition task, so that occluding mouth the loss of information is minor than occluding the eyes. For this reason experiments have been conducted in order to evaluate the degradation of the recognition rate when at least one element of the face is occluded. The test as been done on a subset of 50 subjects from the AR Faces database. Three images have been used with different facial expression: neutral, angry and smile. In the first case only one eye has been occluded in all three images, used as probe set, while non-occluded faces in neutral condition compose the gallery set. In Figure 3 (a) are shown results in terms of CMS when the rank increases from 1 to 30. For neutral images also when one of the eyes is occluded the CMS is 700% just for rank 1, while it drop dawn of about 25-30% for the angry and smile images, because changes in expression introduce a further distortion in the resulting feature vector. In Figure 3 (b) is shown the case, in which both the eyes are occluded. The situation is almost the same for the neutral expression, while at rank 1 the CMS slow down to 50-65% for smile and angry. This means that eyes by itselves provide almost the half of the information needed for recognition. In other words the contribution of the mouth and nose is less significant than that of the eyes, as confirmed also by Figure 3 (c), in which the mouth has been occluded. It can also be observed that for the smiling images the CMS at rank / is higher than for all other types of occlusions, it is because a smiling expression can be just considered as an occlusion, which is amounted to the mouth occlusion in this case. At last Figure 3 (d) shows that also when the nose is occluded a noticeable degradation of the CMS occurs for angry and smiling expression. An interpretation can be that the nose is almost invariant when facial expression changes occur, then occluding nose the recognition task becomes slightly harder. The second interesting aspect to be investigated consists in evaluating how good are the recognition performances when the dimensions of the occluded area increase. In order to test the robustness of ROF in this sense, the same subset of the AR Face database used in the previous experiment has been taken. For each face the neutral expression is used for training the system in the gallery, while neutral, angry and smile images with synthetic occlusion are used for testing in the probe.

170

A. F. Abate, M. Nappi, D. Riccio, and M. Tucci

00 rao-o-o-OHS--^0

I

»

-0



c


,

1 o t 4

*"-4

• ' V'l^

ROF siijiP ROF >«i»W

; - o - ROF !*-•*») i O ROF avi\ i 4 ROF -atnk

(.5 HiBlfc

J »

(a) ! «C

V

/

;'

i * A I, *

»»•/ - * - ROF wimJ I « ROF jmp» ! •0 ROF .m.k' 1 t'